SoylentNews

takyon writes:

Innosilicon graphics cards based on "Fantasy One" GPU feature up to 32GB GDDR6X memory

Today at the "Fantasy One GPU Product Press Conference", Innosilicon, a Chinese company offering graphics and memory solutions unveiled its first discrete GPU.

At the event, Innosilicon revealed its plans for the Fantasy One GPU. This processor is based on Imagination graphics IP (BXT to be specific) which brings tons of innovations to the discrete GPU solutions offered by the Chinese company.

As many as four products have been revealed, including dual-GPU and single-GPU solutions. Type A is a consumer/workstation graphics card featuring a single Fantasy One GPU, which is actually a multi-chip (chiplet) design. Unfortunately, none of the news reports from China on this announcement had the exact configuration of the Fantasy One.

According to the data provided by Innoslicon, this GPU offers up to 5 TFLOPS of single-precision compute power [and a fillrate of 160 GPixel/s]. It is equipped with DisplayPort 1.4, HDMI 2.1, and VGA connectors. This card is to be equipped with up to 16 GB of GDDR6(X) memory across a 128-bit interface. So far, the G6X technology has been exclusive to NVIDIA/Micron Ampere GPUs, but apparently, Innosilicon made a lot of research in PAM4 signaling and was able to unlock up to 19 Gbps memory bandwidth for their GDDR6X implementation. A relatively short memory bus will have its toll on the maximum theoretical bandwidth though, which is to go up to 304 GB/s (so somewhere in between Radeon RX 6700XT and 6600XT).

[...] The Type B [...] is a dual-GPU solution featuring two Fantasy One GPUs connected by an interface known as Innolink. The company claims up to 10 TFLOPS of computing power and 320 GPixel/s fillrates. This card can offer 32 simultaneous 1080p/60FPS streams or 64 streams at 720/30FPS. It is to feature up to 32GB of GDDR6(X) memory but is again limited by dual 128-bit interfaces from each GPU. All these cards feature a PCI-Express 4.0 interface at full X16 width.

When people are willing to pay 2-4x as much for GPUs, competition is inevitable. Intel will be joining the market with "Alchemist" discrete GPUs in Q1/Q2 2022.

Previously: Imagination Announces B-Series GPU IP: Scaling up with Multi-GPU

Original Submission